EXPRESSIVE REINFORCEMENT LEARNING VIA ALGEBRAIC Q-VALUE SCALARIZATION

ant_walking_resized.webm

We can learn controllers such as for Hopper-v4 in 7000 steps!

Screencast.from.2024-09-12.19-47-55.webm

Installation

Via nix

• Install nix from https://determinate.systems/posts/determinate-nix-installer
• Run `nix develop --impure`` in the repo's root, this should take some time but then you will be dropped in a bash shell with everything required to run training

Via Conda

• conda env create --file environment.yml
• conda activate cmorl_env

Run training

python envs/Pendulum/train_pendulum.py, this uses ddpg to train the algorithm and produces automatically checkpoints and logs in the trained folder python envs/Pendulum/test_pendulum.py -lr training